Add support for encoder-only T5 models

[fairydreaming]

This PR adds support for encoder-only T5 models. It also adds support for LLAMA_POOLING_TYPE_NONE in llama-embedding tool for purposes of testing this PR. Fixes #8900.

Model for testing this PR can be downloaded from https://huggingface.co/stabilityai/stable-diffusion-3-medium-diffusers
You have to copy files from text_encoder_3 and tokenizer_3 into a single directory and convert the resulting model with convert_hf_to_gguf.py.

Here's an example transformers script that can be used to print embeddings generated with the original tensorflow model:

from transformers import T5EncoderModel, T5TokenizerFast

id = "stabilityai/stable-diffusion-3-medium-diffusers"

tokenizer_3 = T5TokenizerFast.from_pretrained(
    id,
    subfolder='tokenizer_3'
)

text_inputs = tokenizer_3("translate English to German: The house is wonderful.", return_tensors="pt")
text_input_ids = text_inputs.input_ids
print(text_input_ids)

text_encoder_3 = T5EncoderModel.from_pretrained(
    id, 
    subfolder="text_encoder_3"
)

prompt_embeds = text_encoder_3(text_input_ids)[0]
print(prompt_embeds)

compare the resulting embeddings with the output of the following command:

./llama-embedding -m models/sd3-text-encoder.gguf -p "translate English to German: The house is wonderful." --pooling none --embd-normalize -1

Stable diffusion 3 uses T5TokenizerFast, so there may be some tokenization differences with more complex prompts (llama.cpp Unigram tokenizer is compatible with the "slow" tokenizer class T5Tokenizer).

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High